Land cover classification in an era of big and open data: Optimizing localized implementation and training data selection to improve mapping outcomes

Deriving land cover from remotely sensed data is fundamental to many operational mapping and reporting programs as well as providing core information to support science activities. The ability to generate land cover maps has benefited from free and open access to imagery, as well as increased storage and computational power. The accuracy of the land cover maps is directly linked to the calibration (or training) data used, the predictors and ancillary data included in the classification model, and the implementation of the classification, among other factors (e.g., classification algorithm, land cover heterogeneity). Various means for improving calibration data can be implemented, including using independent datasets to further refine training data prior to mapping. Opportunities also arise from a profusion of possible calibration datasets from pre-existing land cover products (static and time series) and forest inventory maps through to observation from airborne and spaceborne lidar observations. In this research, for the 650 Mha forested ecosystems of Canada, we explored approaches to refine calibration data, integrate novel predictors, and optimize classifier implementation. We refined calibration data using measures of forest vertical structure, integrated novel spatial (via distance-to metrics) model predictors, and implemented a regionalized approach for optimizing training data selection and model-building to ensure local relevance of calibration data and capture of regional variability in land cover conditions. We found that additional vetting of training data involved the removal of 44.7% of erroneous samples (e.g. treed vegetation without vertical structure) from the training pool. Nationally, distance to ephemeral waterbodies was a key predictor of land cover, while the importance of distance to permanent water bodies varied on a regional basis. Regionalization of model implementation ensured that classification models used locally relevant descriptors and resulted in improved classification outcomes (overall accuracy: 77.9% +/- 1.4%) compared to a generalized, national model (70.3% +/- 2.5%). The methodological developments presented herein are portable to other land cover projects, monitoring programs, and remotely sensed data sources. The increasing availability of remotely sensed data for land cover mapping, as well as non-image data for aiding with model development (from calibration data to complementary spatial data layers) provide new opportunities to improve and further automate land cover mapping procedures.

Automated summary

Deriving land cover from remotely sensed data is essential for operational mapping and reporting programs, benefiting from free imagery access and improved technological capabilities. The accuracy of land cover maps depends on calibration data, classification models, and implementation methods.